Let me tell a detective story. Brief one, a bit dull even. Accountants toil daily to match invoices and receipts with transactions. If identifiers like purchase order numbers are missing, the parts then have to be matched by total sum, date and other hints. In psychology, this conscious evaluation of hypotheses and evidence is called ‘system 2 thinking’.
Processes often start from documents like invoices, orders, insurance forms etc. They usually contain the same abstract information but with different layout and keywords. Optical character recognition (OCR) reveals the text in computable form, sprinkled with some errors, but the raw text is still hard to process automatically. A more structured form, like database rows or XML is needed.
A system 2 model of the document helps in correcting the errors, human or computer, and finding the most likely explanation among many possible. An example of a model for invoices: Line items and taxes must add up to the total sum. If not, then an error has been made.
Curious AI’s system 2 solution combines information from the document, a model, and existing databases, giving robust extraction at abstraction level suited for automation. First versions require hand crafted models, but the goal is to automate modeling based on historical data.